Research On Image Enhancement Of Water-sky Scene Under Complex Weather Conditions

As the global environment deteriorates,land resources are gradually being exhausted,and many countries have put their resource strategies into the sea.Therefore,the development of marine equipment is a top priority for all countries.Unmanned Surface Vehicles(USV)are an important part of them and play an important role on both military and civilian.More and more researchers are involved in their related technical research.Among them,the environmental awareness of USV is an important research direction.In view of the problem that the perception of visible light is susceptible to weather,this thesis proposes targeted image enhancement methods for water-sky scenes based on three kinds of weather: sunny,foggy and rainy days.The main contents of this thesis are as follows:Firstly,on sunny days,because of the strong light irradiation,the highlight area generated by the specular reflection of the water surface causes image degradation.Since the traditional highlight removal methods cannot be applied to the water-sky scene,this thesis proposes a highlight detection algorithm based on decision tree and an improved Criminisi repair algorithm.Firstly,by extracting the color features,local structural features and position features of the highlight area,the classification decision tree is trained by the way of EasyEnsemble,and then the water surface to be repaired is obtained.Next,according to the characteristics of the water surface,the improved Criminisi repair algorithm is proposed to compensate the highlight region.It is proved by experiments that the method has strong anti-interference and the image after highlight removal is more natural.Then,in foggy days,the image is degraded because of the effect of fog.For the traditional dehazing method,it is impossible to eliminate the sky effect and to enhance the water surface texture.This thesis proposes a de-fogging method based on local gradient enhancement.Firstly,the accurate atmospheric light A is obtained by the improved quadtree search method,and then the local gradient optimization model is derived according to the characteristics of gradient degradation and sky effect.The coarse transmittance is obtained by solving this optimization model.Next,the guided filtering method is used and the detailed transmittance is obtained.Finally,a fog-free image is obtained according to the atmospheric scattering model.The many experiments prove that our method is practical and can solve the sky effect and image texture degradation.Next,in rainy days,the image is degraded because of raindrops.In view of the fact that the traditional rain removal method cannot remove the rain and fog,this thesis proposes a rain removal method based on generative adversarial network(GAN).Firstly,the development and working principle of the GAN are introduced.Next,the rain-removal network DR-GAN is proposed.The generator G is the multi-task regression network to make the wavelet sub-picture and the dark channel sub-picture of the rain image respectively regress to the wavelet sub-picture and the dark channel sub-picture of the real rain-free image.The discriminator D is the two-class network to distinguish the generated images and real rain-free images.It is proved by experiments that the method has good reinforcement effect and solves the problem of fog in the rain images,further improving the image contrast.Finally,this thesis proposes a weather classification model and analyzes the improvement of the water surface target detection by image enhancement.Firstly,the weather automatic classification model and the data set are proposed.Next,this thesis analyzes the impact of three image enhancement methods on the surface target detection.Finally,it is proved by experiments that image enhancement can increase the accuracy of detection,and the method of this thesis is more effective.